This is the official code for the paper Learning Label Modular Prompts for Text Classification in the Wild (accepted to EMNLP 2022).
Authors: Hailin Chen, Amrita Saha, Shafiq Joty, Steven C.H. Hoi
conda create --name moduPT python=3.6.13
source activate moduPT
conda install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=11.0 -c pytorch
pip install transformers datasets scikit-learn seqeval pickle5 sentencepiece
We use pretrain model T5-large (LM adapted): download t5.1.1.lm100k.large from T5 repo. Put it under ${project_dir}/lm_adapted_models/t5.1.1.lm100k.large
. Then convert it to pytorch checkpoint by
python convert.py --size large --data_root ${project_dir}/lm_adapted_models
ls ${project_dir}/lm_adapted_models/t5.1.1.lm100k.large/pytorch_model.bin
We experiment with three datasets: HuffPost News, FewNERD and FewRel. We randomly sampled fewshots data and split them into multi-stages. You can download the processed data from google drive. Unzip it and put fewNERD
fewrel
huffpost
folders under $project_root/data
directory.
Commands for running training & testing of [modularPrompt
| PromptTuning
| Finetuning
] on HuffPost News:
zsh trainner_hp.sh
zsh trainner_hp_PT.sh
zsh trainner_hp_Finetune.sh
Replace trainner_hp.sh
with trainner_ner.sh
or trainner_re.sh
for other datasets.
You might want to change the following arguments in the above scripts:
Parameters | Description |
---|---|
--lm_adapted_path |
path to lm-adapted T5 checkpoint, e.g. $project_root/lm_adapted_models/t5.1.1.lm100k.large/pytorch_model.bin
|
--cache_dir |
directory to store huggingface model cache |
--mean_prob |
"probability p" in paper section 3.3: the chance to subsample S from |
run_idx |
run id for multi-seed experiments (5 seeds). Default 0 , choose from [0,1,2,3,4] |
The code is released under BSD 3-Clause - see LICENSE.txt
for details.
This code is developed from other open source projects: transformers. We thank the original contributors of these works for open-sourcing their valuable source codes.
If you find our paper or this project helps your research, please kindly consider citing our paper in your publication.
@inproceedings{DBLP:conf/emnlp/ChenSJH22,
author = {Hailin Chen and
Amrita Saha and
Shafiq R. Joty and
Steven C. H. Hoi},
title = {Learning Label Modular Prompts for Text Classification in the Wild},
booktitle = {{EMNLP}},
pages = {1677--1690},
publisher = {Association for Computational Linguistics},
year = {2022}
}